91 / 2025-04-24 12:55:56

Performance of machine learning models and empirical equations on predicting the scour depth around vibrating pipelines

Local scour around a vibrating submarine pipeline threatens its structural stability and
requires accurate predictive models to ensure its safety. This study evaluates the performance of
three standalone Machine Learning (ML) models -- M5 model trees, Adaptive Robust Regression
(ARR), Support Vector Regression (SVR), and an ensemble Gradient Boosting (GB) model on
scour prediction. Their predictive accuracy is compared against four empirical formulas using an
experimental dataset and statistical performance metrics. The results indicate that GB outperforms
all other models, achieving the highest r2 and lowest RMSE and MAPE in the training and testing
phases. M5 and SVR show moderate accuracy, while ARR exhibits the weakest performance.
Empirical equations perform poorly, often significantly overestimating or underestimating scour
depth, demonstrating the limited generalization. Correlation analysis highlights that vibration
amplitude is the dominant factor. These findings emphasize the superiority of ensemble learning
over standalone ML models and empirical equations for improving scour depth predictions.